Intelligent scheduling method for computing power network resources based on large model intention awareness

Abstract

The invention relates to an intelligent computing power network resource scheduling method based on large model intention awareness, belongs to the technical field of automatic configuration of computing power network resources, fully utilizes the advantages of a large language model in the aspect of intention recognition and the advantages of intelligent agent reasoning, planning and the like, and is combined with an external knowledge base to expand the knowledge range of the large language model. Therefore, the answer accuracy is improved, the reasoning burden is reduced, the dynamic updating capability is enhanced, and the adaptability to different fields is improved. Through organic combination with a computing network system, automatic configuration from user natural language description to required resources can be realized, the operation process is simplified, the use threshold is reduced, and more users can easily access and use rich computing power resources. In order to reduce invalid resource configuration in a resource allocation process, a grammar check and parameter check technology based on user description and a template is provided, so that the overall robustness of a program is improved.

Description

Technical Field

[0001] The present invention relates to an intelligent scheduling method for computing power network resources based on large model intention perception, which belongs to the technical field of automatic configuration of computing power network resources. It is committed to solving the problem of intelligent allocation of computing power network resources, and realizes automatic allocation from user natural language description to computing network resources by integrating the powerful capabilities of large language models and intelligent agent technology, thereby improving the intelligence level and efficiency of computing network resource allocation. Background Art

[0002] With the rapid development of science and technology, large language models have flourished under the impetus of the wave of deep learning, and their development has a deep technical accumulation. The early exploration of artificial intelligence and the major breakthroughs in deep learning in the 21st century laid the foundation for large language models. The training of massive data and the use of Transformer architecture have endowed large language models with powerful natural language processing capabilities, enabling them to handle multiple language tasks and understand the context. However, the intelligent agent is different from the concept of intelligent agents in a broad sense. Its core principle is to use large-scale corpora for deep learning, so that the model can understand and generate human language, and then customize personalized AI assistants according to user needs and preferences. Its advantage lies in its flexibility and customization capabilities. It can customize and complete various tasks through instructions, additional knowledge and arbitrary skills. This fusion not only makes up for the limitations of large language models that lack environmental perception and the ability to actively perform tasks, but also greatly improves the intelligence level of the system, enabling intelligent agents to better understand human instructions and perform complex tasks, thus opening a new chapter in the development of intelligence.

[0003] However, in the field of computing networks, intelligent agent technology has not been fully utilized. The deficiencies of traditional computing networks in resource management, elastic expansion, and intelligent scheduling make it difficult to meet the diverse and complex business needs in today's technical environment. In particular, the problem of high user operation thresholds means that users still need to manually select and configure computing network resources, which undoubtedly increases the complexity and difficulty for users who lack professional knowledge. As a result, many users without technical backgrounds face difficulties in configuring and managing computing resources. Users need to have rich technical knowledge to manually select appropriate computing resources (such as CPU, memory, bandwidth, etc.) and configure them reasonably. This process is not only cumbersome, but also prone to configuration errors, further increasing the complexity of resource scheduling. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method for intelligent scheduling of computing power network resources based on large model intent perception. The method makes full use of the advantages of large language models in intent recognition, as well as the advantages of intelligent agent reasoning and planning, and combines with external knowledge bases to expand the knowledge scope of large language models, thereby improving answer accuracy, reducing reasoning burdens, enhancing dynamic update capabilities, and improving adaptability to different fields. Through the organic combination with the computing network system, it can realize the automatic configuration from user natural language description to required resources, simplify the operation process, lower the usage threshold, and enable more users to easily access and use rich computing power resources. In order to reduce invalid resource configuration during resource allocation, syntax checking and parameter checking technology based on user descriptions and templates are proposed to improve the overall robustness of the program.

[0005] The present invention adopts the following technical solution: A method for intelligent scheduling of computing network resources based on large model intention perception includes the following steps: S1: Build multiple intelligent agents with the same functions from different large model service platforms. Each intelligent agent consists of multiple nodes that implement different functions, including intent classification nodes, branch nodes, intent extraction conversion nodes, entity conversion nodes, feasibility analysis nodes, resource classification and template filling nodes, parameter checking nodes, and grammar checking nodes. S2: Design a prompt word template for each node. The prompt word template is a structured input format tailored for large language models to guide the model to generate expected output. S3: Build a standardized computing network resource requirement description template library, attribute semantic mapping table, and intent mapping table; S4: The resource type intent template library, attribute semantic mapping table and intent mapping table are stored in the knowledge base. The knowledge base converts the information into vectors for storage and builds an index structure to quickly retrieve information through a similarity algorithm. S5: Input the user's natural language description into the agent, and after being processed by the intent classification node, branch node, intent extraction conversion node, entity conversion node, feasibility analysis node, resource classification and template filling node, parameter check node and grammar check node, the final user resource demand description template is obtained and configured and issued; S6: During the configuration delivery process, the resources are initially screened according to the resource configuration information in the user resource requirement description template to obtain the candidate resource set T. The scheduling strategy is then matched according to the user resource requirement description template, and the candidate resource set T is further screened and sorted, and finally the appropriate resources are fed back to the user.

[0006] Preferably, in step S1, each node is connected to multiple agents via multiple interfaces to enhance the fault tolerance of the system; The optimal agent is selected through the agent switching algorithm. The agent switching algorithm comprehensively considers three aspects: interface availability, response time and load balancing. The average response time of the interface is calculated by the sliding average method, the interface availability is checked regularly, and a comprehensive load indicator is introduced to quantify the load status of the interface, comprehensively considering the number of requests, processing time and resource utilization; the comprehensive load indicator adopts a dynamic attenuation mechanism to adapt to changes in system load. Finally, the score of each interface is calculated through a comprehensive evaluation indicator. When the interface is unavailable or the response time exceeds the threshold, the available interface with the highest comprehensive evaluation score is selected for switching to ensure the optimal performance of the system.

[0007] Preferably, in step S2, the prompt word template decomposes a total task into multiple subtasks to guide the large language model to execute according to the specified rules, and continuously optimizes the prompt words according to the generated results to meet the current task requirements.

[0008] Preferably, in step S3, a standardized computing network resource requirement description template library, i.e., a resource type intention template library, includes computing network resource configuration information templates such as cloud host configuration information template, storage configuration information template, and network configuration information template, which are described in json format; The process of building a standardized computing network resource requirement description template library is as follows: Conduct demand analysis and research based on the computing network scheduling system to clarify resource allocation requirements; Design the template structure, define the standard attributes of each resource category, and determine the data type and value range; Write resource templates in JSON format and ensure that they comply with the syntax rules of the target platform through syntax checking and logic verification; The resource templates that have passed the verification are stored in the standard computing network resource requirement description template library.

[0009] Preferably, in step S3, the attribute semantic mapping table contains the meanings of the attributes of each field in each standardized computing network resource requirement description template library under a specific computing network platform, so that the open large language model can better learn the knowledge under the specific computing network platform, so as to improve the large language model's ability to understand the standardized computing network resource requirement description template library; The intent mapping table contains as many expressions as possible for each field attribute value, realizing the mapping of user intent to computing network entities and improving the accuracy of large language model generation.

[0010] Preferably, in step S4, the knowledge base parses the uploaded knowledge, and slices the parsed knowledge to form smaller knowledge units that are easier to manage and process, so as to facilitate the vectorization of knowledge; the knowledge units are vectorized using the Embedding vector model and converted into vector form to be stored in the vector database; an index structure is established for the vector database, and the index structure includes an inverted index, a KD tree or an HNSW; and information is quickly retrieved through a similarity algorithm, and the similarity algorithm uses a cosine similarity algorithm: When the user enters a query q, the query is segmented and converted into a word vector to obtain the query vector ; Then, for each document vector stored (Document vector refers to the data uploaded to the knowledge base that is vectorized by the Embedding vector model, converting the document content into a high-dimensional vector form. These vectors capture the semantic information of the document and are stored in the vector database for subsequent efficient retrieval and matching through vector similarity calculations.) Calculate cosine similarity , ,in is the query vector No. k Quantity, is the document vector No. k Quantity, d is the vector dimension, i is an index vector used to distinguish different documents; , , according to the cosine similarity calculation results, several data items that are most similar to the query vector are returned. These data items are the most relevant search results.

[0011] Preferably, in step S5, the intent classification node is completed based on the large language model (LLM), which is used to classify the user intent into two categories: regular questions and answers and configuration information questions and answers according to the user's natural language description, and transmit it to the corresponding process through the branch node. If it is a regular question and answer, the large language model is directly used to answer the user's natural language description; if it is a configuration information question and answer, it is identified as a resource request and enters the subsequent intent extraction conversion node process; The intent extraction conversion node guides the task through the prompt word template, compares the output result of the intent extraction conversion node with the expected result, and optimizes the prompt word template information at the same time, thereby improving the efficiency and accuracy of completing the task, identifying the resource entity and the attribute value of the resource entity in the user's intention, including resource types such as cloud host, memory, network delay, etc., and converting the resource entity into JSON format; When designing the prompt words of the entity conversion node, it is necessary to explicitly require that the node must query the relevant data in the knowledge base when performing tasks to improve the accuracy of the converted data. The entity conversion node further converts the resource type in the preliminary conversion result obtained by the intent extraction conversion node according to the attribute semantic mapping table and intent mapping table in the knowledge base, and obtains the converted json format data in the form of the computing network identifier expression, which is recorded as the computing network entity data.

[0012] Preferably, in step S5, the feasibility analysis node determines whether the computing network platform has sufficient resources for allocation based on the output of the entity conversion node. The feasibility analysis node obtains the resource utilization of the computing network platform by accessing the node in the form of a plug-in by adding an interface for computing network resource query; if any resource in the computing network platform cannot meet the resource configuration required by the user, the user's intention is considered infeasible and feedback is given to the user; otherwise, the user's intention is considered feasible and the next step is performed; when designing the prompt words of the feasibility analysis node, it is necessary to explicitly require the node to use the plug-in to query the remaining resources of the computing network platform before judging whether the computing network platform has sufficient resources for allocation, and then judge whether there are sufficient resources for allocation.

[0013] When the user intention is feasible, the resource classification and template filling node classifies the resource type intention according to the resource type in the computing network entity data, matches the corresponding resource type intention template from the standardized computing network resource demand description template library in the knowledge base according to the classification result, and fills the matching resource type intention template library according to the data in the computing network entity data to obtain the user resource demand description template. When designing the prompt words of the resource classification and template filling node, it is necessary to clearly require that the node must query the relevant data in the knowledge base when performing the task to improve the accuracy of generating the user resource demand description template.

[0014] Preferably, in step S5, the user resource requirement description template obtained by the resource classification and template filling node is subjected to a feasibility check through a parameter check node and a grammar check node. After passing the check, the final user resource requirement description template is obtained. When designing the prompt words of the parameter check node, it is necessary to clearly require how the node performs parameter check. When designing the grammar check node, it is necessary to clearly require the node to use the provided grammar check plug-in to perform grammar check on the user resource requirement description template.

[0015] In order to ensure that the final user intent resource configuration template actually meets the user requirements, it is necessary to perform a modified extraction on the matched resource type intent template, extract the modified content of the matching template, and then compare the modified content with the information of the entity conversion node to ensure that the modified content is consistent with the user requirements. The parameter check node compares the output of the resource classification and template filling node with the corresponding resource type intent template in the knowledge base, extracts the different content between the two, that is, the content of the resource classification and template filling node that modifies the resource type intent template, denoted as U1, and then compares U1 with the output of the entity conversion node. If they are the same, it is considered to meet the requirements, otherwise it does not meet the requirements, and the content of the resource classification and template filling node is modified according to the output of the entity conversion node, and output according to the modified content; Although the large language model can generate grammatically fluent text, due to the differences in the requirements of different computing network platforms for configuration information, such as naming format, attribute data type, value range, etc., the resource configuration template generated by the large language model cannot be guaranteed to be completely correct under specific grammatical rules. The large language model may generate text that does not conform to the grammar specification (such as missing necessary fields, format errors, key-value pair mismatches, etc.). Therefore, the purpose of grammar checking is to verify whether the configuration generated by the large language model meets the grammatical requirements of the target platform, to ensure that it can be correctly parsed and executed by the scheduling system, and to avoid resource configuration failures or system failures due to grammatical errors. When formulating grammar rules, the grammatical requirements and template format of the target platform should be strictly followed to ensure that the rules can effectively match the actual needs of the target platform.

[0016] According to the requirements of the computing network platform, grammar rules are pre-formulated. The grammar checking node performs grammar analysis on the generated user resource demand description template according to the pre-formulated grammar rules, checks whether the user resource demand description template generated by the intelligent agent has errors that violate grammar rules, and captures error information; If there is no syntax error, a final user resource requirement description template is generated; if there is a syntax error, the user resource requirement description template is modified according to the error information until there is no syntax error.

[0017] For any details not provided in the present invention, please refer to the prior art.

[0018] The beneficial effects of the present invention are: The present invention integrates the intention perception of large language models and intelligent agent technology to realize an intelligent and efficient process from user natural language description to automatic resource allocation. By building intelligent agents, multiple interfaces and intelligent agent switching algorithms, the system is guaranteed to be stable and efficient, the knowledge base is used to improve the knowledge level of the model, and the final user resource demand description template is ensured to be accurate and reasonable through intention perception conversion and feasibility testing. A friendly interface is provided to lower the threshold for user use, improve the usability of the computing network platform, and promote the intelligent development of computing network resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0020] Figure 1 It is a schematic diagram of the overall structure of the system of the present invention; Figure 2 Construct an overall framework diagram for the intelligent agent of the present invention; Figure 3 It is a flow chart of parameter checking and grammar checking of the present invention; Figure 4 It is a flow chart of the intelligent scheduling method of computing power network resources based on large model intention perception of the present invention; Figure 5 It is a schematic diagram of the workflow of each node of the intelligent agent of the present invention; Figure 6 Schematic diagram of intelligent agent switching of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the drawings in the implementation of this specification, but are not limited to this. Anything not fully described in the present invention shall be based on the conventional technology in the art.

[0022] Terminology explanation: 1. Large Language Model: Large Language Model (LLM) is an artificial intelligence model composed of billions or even trillions of parameters, which is trained with a large amount of text data. These models can generate grammatically correct, semantically coherent, and contextually reasonable natural language texts, and can perform tasks such as text generation, translation, dialogue, question answering, and code generation.

[0023] 2. Agent: Different from the general concept of agent, it is derived from the large language model and can customize its own AI assistant according to the needs and preferences of users. Its advantages lie in its flexibility and customization ability, which can customize and complete various tasks through instructions, additional knowledge and arbitrary skills. Its core principle is to use large-scale corpus for deep learning, so that the model can understand and generate human language, and then customize personalized AI assistant according to the needs and preferences of users.

[0024] 3. Intent perception: Intent perception means that the system understands the user's goals, needs or purposes by analyzing the user's input and converting it into structured data that the machine can understand and process. In simple terms, it is to identify what the user wants to do or want by parsing the user's language.

[0025] 4. Intent perception with large language models: Large language models, such as the GPT series (including GPT-3, GPT-4, etc.), ChatGLM series, and BERT, can identify user intent through contextual understanding, syntactic analysis, and semantic reasoning. Compared with traditional intent perception methods, large language models are generally more flexible and powerful, and can handle complex language expressions, ambiguous intents, and cross-domain tasks.

[0026] 5. Standardized computing network resource requirement description template library: It is a resource configuration template for creating various cloud resources under a specific computing network platform and can be called and executed by the scheduling system. It contains the configuration information required by users to create cloud resources, such as resource type, number of CPU cores, memory size, network bandwidth, network latency, disk capacity, etc.

[0027] 6. Attribute semantic mapping table: used to explain the standardized computing network resource demand template. Explain the meaning of each field in the standardized computing network resource demand template. Help the big language model understand the representation of the element in a specific computing network platform to ensure the consistency, validity and standardization of the data.

[0028] 7. Intent mapping table: It can convert different expressions of the same resource into specific values ​​in the computing network resource template. Since users express their intentions in a variety of ways, it is not feasible to directly use these intentions as the values ​​of the resource template. Therefore, the purpose of the intent mapping table is to collect and organize as many expressions of each attribute value as possible, so as to accurately convert the user's description into the corresponding expression in the computing network resource template.

[0029] 8. Knowledge base: In the field of artificial intelligence (AI), a knowledge base refers to a system for storing, organizing and processing knowledge. It contains facts, rules, concepts, relationships and reasoning mechanisms about a certain field. Its purpose is to support the understanding, reasoning, decision-making and learning of large models.

[0030] 9. Prompt word template: In artificial intelligence (especially natural language processing, NLP), a preset text or structure is used to guide and trigger the model to generate a specific type of output. Prompt word templates usually include predefined text structures, questions or instructions. By inputting specific prompt words, the AI ​​model can understand and generate expected answers or behaviors. It is used to ensure that the model can efficiently generate accurate, relevant and high-quality responses. By using templates, the ambiguity of the model in generating content can be reduced and the output quality of the model can be improved, especially when facing complex problems or multi-round conversations.

[0031] 10. Syntax check: Syntax check refers specifically to syntax check in programming. In programming, a set of strict grammar rules are formulated in advance, and the code is parsed according to the grammar rules to check whether there are errors that violate the grammar rules, ensuring that the code can be correctly parsed and executed. In the present invention, syntax check is mainly to check the data in json format.

[0032] 11. Parameter check: Verify the value of a property element to ensure that it meets the user's preset requirements and rules, so as to ensure that the program will not make errors during processing due to invalid or non-compliant input. Parameter check is not only the basis for ensuring function implementation, but also a key link to improve system stability, usability and user experience.

[0033] 12. Intent classification node, branch node, intent extraction conversion node, entity conversion node, feasibility analysis node, resource classification and template filling node, parameter checking node, grammar checking node: These nodes are based on the large language model and combined with external tools to complete the corresponding tasks.

[0034] Example 1 A method for intelligent scheduling of computing network resources based on large model intention perception includes the following steps: S1: Build multiple intelligent agents with the same functions from different large model service platforms, such as Wenxin Yiyan, Zhipu Qingyan, Tongyi Qianwen and other platforms. Each intelligent agent is composed of multiple nodes that implement different functions, including intent classification nodes, branch nodes, intent extraction conversion nodes, entity conversion nodes, feasibility analysis nodes, resource classification and template filling nodes, parameter checking nodes and grammar checking nodes. These nodes cooperate with each other to complete tasks, such as Figure 2 As shown; S2: Design a prompt word template for each node. The prompt word template is a structured input format tailored for large language models to guide the model to generate expected output. S3: Build a standardized computing network resource requirement description template library, attribute semantic mapping table, and intent mapping table; S4: Store the resource type intent template library, attribute semantic mapping table and intent mapping table in the knowledge base. The knowledge base is located on the model service platform of each intelligent agent. The knowledge base converts information into vectors for storage and builds an index structure to quickly retrieve information through a similarity algorithm. S5: Input the user's natural language description into the agent, and after being processed by the intent classification node, branch node, intent extraction conversion node, entity conversion node, feasibility analysis node, resource classification and template filling node, parameter check node and grammar check node, the final user resource demand description template is obtained and configured and issued, such as Figure 4 and Figure 5 As shown; S6: During the configuration delivery process, the computing network scheduling system performs preliminary resource screening according to the resource configuration information in the user resource demand description template (according to the resource requirements in the user resource demand description template, such as CPU quantity, memory size, disk size and other resource information, first screen out all computing network nodes that can meet the user's requirements in the entire computing network node, that is, the screened computing network nodes should meet the remaining resource amount of the computing network node should be greater than the resource amount required by the user), obtain the candidate resource set T, and then match the scheduling strategy according to the attribute value of the strategy_name field in the user resource demand description template, further screen and sort from the candidate resource set T, and finally feed back the appropriate resources to the user.

[0035] After completing the syntax and parameter verification of the standard user resource requirement description template, the scheduling system will extract the final user resource requirement description template and automatically perform intelligent scheduling of resources to ensure optimal configuration and efficient utilization.

[0036] Exemplarily, suppose the candidate resource set initially screened according to the resource configuration information is T={T1, T2, T3, T4}, when the template field strategy_name: price_priot means price priority, then the candidate resource set T is sorted from low to high in price to obtain TP={T2, T1, T4, T3}, and finally the cheapest resource T2 is fed back to the user.

[0037] Example 2 A method for intelligent scheduling of computing network resources based on large model intention perception is shown in Example 1, except that in step S1, the construction of the intelligent agent comprehensively considers factors such as the complexity of the task and the accuracy of the large language model generation, divides the intelligent agent into multiple nodes, and divides a complete task into several subtasks to complete, which can effectively reduce the complexity of the task. At the same time, the combination of external knowledge base, prompt word engineering and specific plug-ins can effectively improve the accuracy of the large language model generation and the efficiency of task completion.

[0038] In order to avoid the program not being able to run normally due to the failure of a single large model interface, each node is connected to multiple agents through multiple interfaces to enhance the fault tolerance of the system and avoid single point failure; The optimal agent is selected through the agent switching algorithm to achieve fast switching of agent interfaces, such as Figure 6 As shown. The agent switching algorithm comprehensively considers three aspects: interface availability, response time and load balancing. The average response time of the interface is calculated by the sliding average method, the interface availability is regularly checked, and the comprehensive load index is introduced to quantify the load status of the interface, taking into account the number of requests, processing time and resource utilization; the comprehensive load index adopts a dynamic attenuation mechanism to adapt to changes in system load. Finally, the score of each interface is calculated through a comprehensive evaluation index. When the interface is unavailable or the response time exceeds the threshold, the available interface with the highest comprehensive evaluation score is selected for switching to ensure the optimal performance of the system, so as to ensure that the agent interface can be quickly switched when an interface failure occurs, and to ensure the stability of the program.

[0039] Specific: Response time RT i = T end - T start ,in T start is the timestamp when the request starts to be sent, T end is the timestamp when the interface response is completed. RT i is the response time of the interface, i Indicates i Agent interface; In order to evaluate the response time more accurately, the moving average method is used to calculate the average response time ART i ,set up RT i , k Indicates i Agent Interface k The response time of each request, mThe window size for calculating the average response time (e.g. m =10), then: ,in n The sequence number of the current request; Interface availability: Send heartbeat requests regularly (e.g., every 10 seconds) to each agent interface and record the accessibility status of the interface. A i , the initial value is 1, indicating that it is available. n times (such as n =3) When the heartbeat request does not receive a response, A i Set to 0, indicating that it is unavailable. i Indicates i An intelligent agent interface.

[0040] Load Balancing: (1) Load index quantification model: Introduce a comprehensive load index for each agent interface LI i ( i Indicates i This indicator comprehensively considers multiple factors such as the number of requests, average request processing time, resource utilization, etc., in order to more accurately reflect the interface load status. LI i =0.

[0041] LI i = α*RQ i +β*BRT i +γ*RU i in, α , β , γ are weight coefficients, representing the number of requests ( RQ i ), average request processing time ( BRT i ), resource utilization ( RU i ) in the load index, and α + β + γ = 1, to ensure that each agent interface is as overloaded as possible, α A larger weight can be taken, such as α =0.5, β =0.2, γ =0.3.

[0042] Number of requests RQ i : Whenever the agent interface processes a request, RQ i Add 1 to directly reflect the request traffic of the interface; Average request processing time BRT i : Record the processing start time of each request t start and end time t end , calculate the processing time of each request BRT i , k = t end - t start ( k Indicates k Requests). Then the exponentially weighted moving average (EWMA) method is used to calculate the average request processing time. The formula is: BRT i =λ* BRT i , k + (1-λ)* BRT i , k-1 Among them, λ is the smoothing factor, and its value range is 0<λ<1. The smoothing factor λ determines the influence of historical data on the current average value in the exponentially weighted moving average method (EWMA). Considering the situation of more user requests, in order to ensure the stable operation of the system, a smaller λ value can be selected, such as between 0.1 and 0.3, to prevent the average value from being affected by short-term fluctuations.

[0043] BRT i , k-1 The average processing time calculated last time. This method can give a higher weight to the processing time of recent requests and reflect changes in interface performance more promptly.

[0044] Resource Utilization RU i :Use system monitoring tools to obtain the percentage of CPU, memory, network bandwidth and other resources used by the agent interface during the request processing. usage is the CPU utilization, Memory usage is memory utilization, Bandwidth usage is the network bandwidth utilization, then: RU i =w1*CPUusage +w2*Memory usage +w3*Bandwidth usage , where w1, w2, and w3 are weight coefficients of the corresponding resource utilization, and w1 + w2 + w3 = 1. For example, if the CPU has the most important impact on the system, followed by memory, w1, w2, and w3 can be set to 0.5, 0.4, and 0.1 respectively, and can be set according to the importance of each resource on performance in the actual system.

[0045] (2) Dynamic load index attenuation mechanism: The load index is attenuated regularly (e.g., every 60 seconds) to adapt to the dynamic changes in the system load and avoid excessive influence of historical load data on current decisions. Assume that the attenuation factor is δ (0<δ<1, e.g., δ=0.9), and the calculation formula for the attenuated load index is: LI inew = LI i *δ.

[0046] Finally, the comprehensive evaluation index is calculated: Calculate comprehensive evaluation indicators for each agent interface SI i , the formula is as follows: SI i =m1* A i +m2*(1 / ART i )+m3*(1 / LI inew ) Among them, m1, m2, and m3 are weight coefficients, which respectively represent the importance of interface availability, response time, and load balancing in the comprehensive evaluation, and m1+ m2+ m3=1. The weights m1, m2, and m3 can be adjusted regularly or based on factors such as load and response time according to actual conditions. For example, m1=0.3, m2=0.4, and m3=0.3 can be set.

[0047] Switching condition: During program operation, if the interface is unavailable, A i =0, average interface response time ART i Exceeding the set threshold ART th Then the interface is switched. First, select the available nodes, namely A i =1, recorded as AM={p1,p2,…,pi}, where p1,p2,…pi represent the first, second,…ith agent interface respectively, and then these agent nodes are evaluated comprehensively. SI i Calculation of SLM ={SL1,SL2…SLi}, SLM Represents the result of the comprehensive evaluation index corresponding to each agent node. The agent node with the largest comprehensive evaluation index is selected as the target switching node.

[0048] Example 3 A method for intelligent scheduling of computing power network resources based on large model intention perception, as shown in Example 2, the difference is that in step S2, the prompt word template decomposes a total task into multiple subtasks to guide the large language model to execute according to the specified rules, and continuously optimizes the prompt words according to the generated results to meet the current task requirements.

[0049] The prompt word template provides clear context, task guidance and format requirements to help the model understand user needs more accurately, control the format and style of generated content, improve the accuracy of generation and the efficiency of task execution, reduce generation errors and optimize user experience. The prompt word template enables the large language model to complete the specified task according to the predetermined requirements and generate a resource demand description template that is more in line with the user and the computing network platform.

[0050] Example 4 A method for intelligent scheduling of computing network resources based on large model intent perception, as shown in Example 3, the difference is that due to the diversity of computing network resources, including but not limited to cloud hosts, storage, networks and other resources, it is necessary to build multiple standardized computing network resource demand description templates that match them, which are used to classify resource intents and match the corresponding resource type intent templates, and fill in the necessary information of the matched resource type intent templates. The construction of this template should include the configuration information required by the resource scheduling system under the computing network platform, in order to enable the scheduled system to effectively schedule resources according to the template, and then the scheduling system will screen out suitable resources according to the configuration information in the template and feedback to the user.

[0051] In step S3, a standardized computing network resource requirement description template library, i.e., a resource type intention template library, is used to describe computing network resource configuration information templates such as cloud host configuration information template, storage configuration information template, and network configuration information template in json format; These resource type intent templates contain the configuration information required to create this type of resource. In the resource type intent template library, each resource category template presets the standard attribute values ​​of this category of resources. This means that as long as the computing network resources are sufficient, the scheduling system can perform efficient resource allocation based on these templates.

[0052] The process of building a standardized computing network resource requirement description template library is as follows: Conduct demand analysis and research based on the computing network scheduling system to clarify resource configuration requirements, including specific requirements for cloud hosts, storage, networks and other resources; Design the template structure, define the standard attributes of each resource category (such as CPU, memory, storage capacity, etc.), and determine the data type and value range; Write resource templates in JSON format and ensure that they comply with the syntax rules of the target platform through syntax checking and logic verification; The resource templates that have passed verification are stored in the standardized computing network resource requirement description template library and integrated with the scheduling system to ensure that the scheduling system can correctly parse and use these templates.

[0053] Exemplarily, the cloud host configuration information template is: {"source_type":"cloud_host","auto_start":true","sn_node":"jn_1","cloud_type": "public","strategy_name": "price_priority","bill_type":"postpaid","auto_renew": true,"use_time_stage": "allday", "opera_system": "windows","description":"null", "hypervisor":"openstack","cloud_count":1,"disks":[{"disk_type":"sys", "index":0,"backend":"SSD","disk_size":1024,"image_id":"null", "medium":"rotate"}],"nets":[{"nets_exit":false}],"network_parameter":[{"width_value": "1G / S","delay_value":"1ms","jitter_value": "50ms"}], "vcpu_count":2,"vmem_size":4096,"project_id":"null","deploy_telegraf":true, "request_id":"null"}This resource template represents the template required by the computing network scheduling system when creating a cloud host. When constructing a resource template, you must strictly follow the requirements of the computing network platform scheduling system to ensure that the constructed resource template can be effectively called by the scheduling system and successfully create the corresponding cloud resources.

[0054] Attribute semantic mapping table: Since large language models are trained on large data sets, there are knowledge blind spots for unknown knowledge. Therefore, directly allowing the large language model to fill in the template according to the user's intention and in combination with the various resource type intention templates in the constructed standardized computing network resource requirement description template library may result in filling errors because the large model cannot directly and accurately identify the meaning represented by each attribute in the template. Therefore, in order to enable the open large language model to understand the information in the standardized computing network resource requirement description template, multiple attribute semantic mapping tables that match the standardized computing network resource requirement description template are constructed. The attribute semantic mapping table contains the meaning represented by each field attribute in the standardized computing network resource requirement description template library under a specific computing network platform, so that the open large language model can better learn the knowledge under the specific computing network platform, so as to improve the large language model's understanding of the standardized computing network resource requirement description template library, so that the large language model can accurately map to the corresponding attribute field according to the user's intention; The attribute semantic mapping table is constructed to match the standardized computing network resource requirement description template, which is used to explain the attributes contained in the resource template and clarify the specific meaning of these attributes in the computing network platform.

[0055] In schematic form, taking the standard computing network resource requirement description template for creating a cloud host as an example, the corresponding attribute semantic mapping table is as follows: Table 1 Attribute semantic mapping table

[0056] The first column of Table 1 shows the attribute names in the resource template, while the second column details the specific meanings of these attributes in the computing network platform.

[0057] Intent mapping table: Since different users have different ways of expressing the same resource, and user intent cannot be directly used as the value in the resource template, it is necessary to convert the user intent into the corresponding computing network expression. The intent mapping table contains as many expressions as possible for each field attribute value, realizing the mapping of user intent to computing network entities and improving the accuracy of large language model generation.

[0058] Exemplary: The following table is the corresponding intent mapping table; Table 2 Intent mapping table

[0059] Example 5 A method for intelligent scheduling of computing network resources based on large model intention perception, as shown in Example 4, is different in that in step S4, the knowledge base provides sufficient external knowledge for the large language model, thereby enhancing the model's ability to understand template configuration information and improving the accuracy of its generated content. Since the knowledge base can convert corpus information into vectors for storage and build an index structure, and quickly retrieve information through a similarity algorithm, it can effectively improve the accuracy and speed of retrieval compared to traditional databases.

[0060] The knowledge base parses the uploaded knowledge, decomposes it from the original format into meaningful components, and slices the parsed knowledge, dividing the knowledge content according to certain rules or standards to form smaller, easier to manage and process knowledge units to facilitate the vectorization of knowledge; the knowledge units are vectorized using the Embedding vector model and converted into vector form to be stored in the vector database, and can better represent the relationship between knowledge units; the vector database is indexed to speed up similarity search. The index structure includes inverted index, KD tree or HNSW (hierarchical-navigable-small-world-graph), etc.; the similarity algorithm is used to quickly retrieve information, convert the user's query into vector form, and quickly search from the vector database through the index structure and similarity algorithm, and return the data results most similar to the query vector. The similarity algorithm uses the cosine similarity algorithm: When the user enters a query q, the query is segmented and converted into a word vector to obtain the query vector ; Then, for each document vector stored , calculate the cosine similarity , ,in is the query vector No. k Quantity, is the document vector No. kQuantity, d is the vector dimension, i is an index vector used to distinguish different documents; , , according to the cosine similarity calculation results, several data items that are most similar to the query vector are returned. These data items are the most relevant search results.

[0061] Example 6 A method for intelligent scheduling of computing network resources based on large model intention perception, as shown in Example 5, the difference is that in step S5, the intention classification node is completed based on the large language model (LLM), which is used to realize that the user intention is divided into two categories according to the user's natural language description: general questions and answers and configuration information questions and answers, and is transmitted to the corresponding process through the branch node. If it is a general question and answer, the large language model is directly used to answer the user's natural language description; if it is a configuration information question and answer, it is identified as a resource request and enters the subsequent intention extraction conversion node process; To classify user natural language descriptions into general questions and configuration information questions and answers, and to determine whether they are resource requests, we can distinguish them by analyzing the keywords, verbs, and context in the input. General questions and answers usually involve queries for information, such as "What's the weather like today" or "How do I set my account password?" The core of these questions is to obtain information or operational guidance. In contrast, configuration information questions and answers focus on the creation, modification, or query of resources, and usually contain resource-related keywords (such as "CPU," "memory," "storage," etc.) and descriptions related to quantity and units (such as "4 cores," "8GB"). In addition, verbs such as "create," "modify," and "extend" usually indicate that the user is requesting a configuration operation.

[0062] The branch node selects the configuration information process or the daily question and answer process based on the output of the intent classification node. The branch node will select the "configuration information question and answer" process and execute the relevant process.

[0063] The intent extraction conversion node guides the task through the prompt word template, compares the output result of the intent extraction conversion node with the expected result, and optimizes the prompt word template information at the same time, thereby improving the efficiency and accuracy of completing the task, identifying the resource entity and the attribute value of the resource entity in the user's intention, including resource types such as cloud host, memory, network delay, etc., and converting the resource entity into JSON format; The entity conversion node further converts the resource type in the preliminary conversion result obtained by the attribute semantic mapping table and intent mapping table in the knowledge base and the intent extraction conversion node in the form of the computing network identifier expression to obtain the converted json format data, which is recorded as the computing network entity data.

[0064] The output of the intent extraction conversion node is used as the input of the model. Unlike the intent extraction conversion node, the entity conversion node uses a knowledge base tool and takes the attribute semantic mapping table and intent mapping table of various resources as the external knowledge source of the node. Then, according to the resource type in the result obtained by the intent extraction conversion node, the corresponding attribute semantic mapping table and intent mapping table are matched from the knowledge base, thereby providing the node with rich contextual information and accurate entity recognition capabilities, improving the accuracy of the node's conversion from user intent entities to computing network attributes.

[0065] For example, the prompt word for extracting the conversion node is "##Role You are an expert in resource intent extraction and are good at accurately converting the extracted resource information into the corresponding JSON data format.

[0066] ##Require 1. You need to accurately identify the configuration information and resource type required by the user.

[0067] 2. Convert the configuration information into json format.

[0068] 3. Only convert the configuration proposed by the user into json format, and do not add any that are not mentioned.

[0069] 4. You need to consider the changes in user needs based on the conversation information {{historical conversation}} and the current conversation {{user input-conversation}}.

[0070] 5. Only the converted json content needs to be output, and the rest does not need to be output.

[0071] ##Example For example, the user's natural language description is "Create a cloud host with a CPU of 2 cores, a memory of 4GB, and a disk size of 40GB." Convert to json{ Resource type: cloud host, Number of cloud hosts: 1, Number of CPU cores: 2, Memory size: 4GB, Disk size: 40GB }” Through this prompt word, the model will extract the user's intention and convert it into JSON data about the entity. The output of the intention extraction conversion node is { Resource Type: Cloud Host, “Number of cloud hosts”: 1, "Number of CPU cores": 4, "Memory size": 8GB, Disk size: 20GB } The prompt word for the entity conversion node is "##Role You are a JSON data conversion expert who can convert known JSON data according to the mapping relationship.

[0072] ##Require 1. You need to {{extract configuration information into json}} based on the obtained json data. You must use the knowledge base tool "{attribute semantic mapping table}", then accurately find the mapping field corresponding to the attribute and convert it. You must use the knowledge base tool "{intention mapping table}", then accurately find the mapping field corresponding to the attribute value and convert it, and finally output the converted json data.

[0073] 2. Only the converted json data needs to be output, and the rest of the content does not need to be output.

[0074] ##example For example, the initial json data is { "Resource Type": Cloud Host, "Number of cloud hosts": 10, "cpu cores": 8, "Memory size": 32GB } The converted json data is: { "source_type": "cloud_host", "cloud_count":10, "vcpu_count":8, "vmem_size":32768 } Finally, the converted json data is output.

[0075] ” Through the prompt word instruction, the entity conversion node must first search the knowledge base to find the corresponding mapping field before performing the conversion, and improve the accuracy of the node's answer by continuously optimizing the prompt word instruction (according to the output result of the node, judge whether the node's output is stable and meets the task requirements. If not, optimize the prompt word according to the output result to make the node's output more accurate, stable and meet the expected goals, thereby improving the reliability of the entire system). This process effectively improves the accuracy of the conversion and reduces the risk of "hallucination" of the generated results. The output of the entity conversion node is: { "source_type":"cloud_host" "cloud_count": 1, "vcpu_count": 4, “vmem_size”: 8GB, “disk_size”: 20GB } The feasibility analysis node determines whether the computing network platform has sufficient resources for allocation based on the output of the entity conversion node. The feasibility analysis node obtains the resource utilization of the computing network platform by adding the computing network resource query interface to the node in the form of a plug-in; suppose that the remaining CPU amount in the computing network node is scpu_count, the remaining memory amount is smem_size, and the remaining disk amount is sdisk_size queried through the interface plug-in. If the resource configuration information required by the user is that the number of CPUs is vcpu_count, the number of memories is vmem_size, and the disk size is disk_size, then compare the size between scpu_count and vcpu_count, the size between smem_size and vmem_size, and the size between sdisk_size and disk_size. If any resource in the computing network platform cannot meet the resource configuration required by the user, the user's intention is considered infeasible and feedback is given to the user; otherwise, the user's intention is considered feasible and the next step is taken; When the user intention is feasible, the resource classification and template filling node classifies the resource type intention according to the resource type in the computing network entity data, matches the corresponding resource type intention template from the standardized computing network resource demand description template library in the knowledge base according to the classification result, and fills the matching resource type intention template library according to the data in the computing network entity data to obtain the user resource demand description template.

[0076] The knowledge base of resource classification and template filling nodes contains multiple resource type intention templates and attribute semantic mapping tables for different resource types. The resource type is first determined based on the attribute value of the source_type field in the feasibility analysis node output through the prompt word template, and then the corresponding resource type intention template is selected from the knowledge base and filled in according to the data of the feasibility analysis node (each field in the resource type intention template has a default attribute value, and the individual fields in the resource type intention template are modified to the values ​​of the fields proposed in the feasibility analysis node.), and the user resource demand description template is obtained.

[0077] For example, the feasibility analysis output is: { "source_type":"cloud_host" "cloud_count": 1, "vcpu_count": 4, “vmem_size”: 8GB, “disk_size”: 20GB } Resource classification and template filling node prompt words “ ##Role You are an expert in resource classification and template filling ##Require 1. Determine the resource type based on the value of the attribute in "source_type": "cloud_host" and query the attribute semantic mapping table in the knowledge base.

[0078] 2. Based on the judgment result, the resource type intention template in the corresponding knowledge base must be matched.

[0079] 3. After the matching is completed, modify the matching template content according to the input configuration information.

[0080] 4. For requirements not mentioned by the user, the fields in the template use default values.

[0081] 5. Finally, output the modified template and do not output unnecessary content.

[0082] ” The resource classification and template filling node searches the attribute semantic mapping table according to the value of source_type cloud_host, determines it as the cloud host type, then searches the knowledge base resource type intent template and matches it to the cloud host template, and fills it. {"source_type":"cloud_host","auto_start":true,"generate_name":"123","sn_node":"jn_1","cloud_type": "public","strategy_name": "price_priority","bill_type": "postpaid","auto_renew": true,"use_time_stage": "allday", "opera_system": "windows","description":"null", "hypervisor":"openstack","cloud_count":1,"disks":[{"disk_type":"sys", "index":0,"backend":"SSD","disk_size":20,"image_id":"null", "medium":"rotate"}],"nets":[{"nets_exit":false}],"network_parameter":[{"width_value": "1G / S","delay_value":"1ms","jitter_value": "50ms"}], "prefer_region":"null","vcpu_count":4,"vmem_size":8,"project_id":"null","sku":"null","duration":"allday","deploy_telegraf":true, "request_id":"null"}Get the user resource requirement description template.

[0083] Example 7 A method for intelligent scheduling of computing network resources based on large model intention perception, as shown in Example 6, the difference is that in step S5, the user resource demand description template obtained by resource classification and template filling nodes is subjected to feasibility check through parameter check nodes and grammar check nodes. After the check passes, the final user resource demand description template is obtained.

[0084] Preferably, the parameter checking node compares the output of the resource classification and template filling node with the corresponding resource type intent template in the knowledge base, extracts the different contents between the two, i.e., the contents modified by the resource classification and template filling node to the resource type intent template, recorded as U1, and then compares U1 with the output of the entity conversion node. If they are the same, it is considered to meet the requirements, otherwise it does not meet the requirements, and the contents of the resource classification and template filling node are modified according to the output of the entity conversion node, and output according to the modified contents; According to the requirements of the computing network platform, grammar rules and processes are pre-established and plug-ins are provided for the grammar check node to use. The grammar check node performs grammar analysis on the generated user resource demand description template according to the pre-established grammar rules and the json.loads() function in the json module to check whether there are any errors that violate the grammar rules in the user resource demand description template generated by the agent, such as Figure 3 As shown, it includes basic grammatical rules, data types, characters, spaces and line breaks, values, etc., and captures error information; if the generated user resource demand description template lacks necessary fields, attribute values ​​are out of range, field naming is incorrect, etc., if there is a grammatical error, the node will modify the user resource demand description template according to the error information until there is no grammatical error, so as to avoid resource configuration failure or system failure caused by grammatical errors, and ensure that the user resource demand description template can be correctly parsed and executed. In the present invention, grammar checking is mainly to check data in json format. When formulating grammar rules for grammar checking, the grammar requirements and template format of the target platform should be followed, so that the formulated grammar rules are actually in line with the target platform.

[0085] Example 8 A computing network resource intelligent scheduling method based on large model intention perception is based on the following modules: intelligent agent construction module, intelligent agent encapsulation module, intention perception conversion module, knowledge base module, verification module, and business delivery module. Figure 1 shown.

[0086] The intelligent agent construction module builds intelligent agents with the same functions on different large model service platforms. Each intelligent agent is composed of multiple nodes with different functions, which work together to complete a task.

[0087] The intelligent agent encapsulation module includes a large language model intelligent agent interface that integrates multiple different large model service platforms to avoid single point failures causing the entire system to malfunction, and designs an intelligent agent switching algorithm to achieve fast switching between multiple interfaces.

[0088] Knowledge base module, because in the field of computing power network, different data centers lack unified standards for the description and measurement of computing power resources, which makes the computing power and network resource configuration of a specific platform, even if the large language model has been trained on a large-scale data set, there is still a lack of knowledge. In view of this, the present invention constructs a computing network platform resource information library composed of a standardized computing network resource demand description template library and corresponding attribute semantic mapping tables and intent mapping tables, and stores it in the knowledge base. The construction of this knowledge base is intended to provide a strong support for the large language model to learn efficiently in a specific environment. When facing user questions, the large language model can accurately search and learn in the knowledge base, so as to master the detailed knowledge of the specific platform, and then realize the accurate conversion from user business intentions to specific platform resource attributes, effectively ensuring that business needs can be comprehensively and accurately identified and converted, greatly improving the intelligence level and accuracy of computing power network resource configuration.

[0089] The core function of the intention-aware conversion module is to accurately convert user intentions into user resource demand description templates that can be directly executed by the resource scheduling system. Specifically, it covers the following key components: (1) Construction of prompt word templates, which assists the model nodes in the intelligent agent to understand the task requirements, clarify the context, adjust the answer style, improve execution efficiency, and guide the model to generate more accurate, relevant and useful outputs. (2) Extraction and conversion of user intention entities and entity conversion, which uses prompt word templates to guide model nodes to extract user intention entities and convert them into specific format data based on the knowledge base to achieve preliminary structuring of user intentions. (3) Feasibility analysis, which obtains the remaining resource information of the computing network platform by connecting the computing network resource query interface plug-in to the model node, and then judges the feasibility of the user intention based on the remaining resource information of the computing network platform and the user resource information. (4) Resource classification and template filling, after feasibility analysis, if the user intention is feasible, the resource type intention is classified, and the corresponding resource type intention template is matched from the knowledge base according to the classification result, and the template is filled with necessary information based on the data obtained in the process of extraction and conversion of user intention entities and entity conversion.

[0090] The verification module is responsible for checking whether the standard user resource demand description template generated by the agent is correct, including syntax checking and parameter checking. Due to the "hallucination" problem of the large language model, this problem may generate misleading or erroneous information when answering user questions, which will cause the generated standard user resource demand description template to deviate or be wrong, and then cause the template to fail to meet the requirements of the resource scheduling system of the computing network platform. Ultimately, users cannot smoothly obtain the required computing resources, which greatly affects the user experience and the service quality of the system. Syntax checking checks whether the content of the user resource demand description template conforms to the predetermined grammatical rules by defining grammatical rules. Parameter checking is to ensure that the configuration information of the user resource demand description template is consistent with user requirements. This method can effectively discover potential grammatical errors and verify the rationality of parameter configuration, thereby improving the overall robustness of the template and reducing unexpected errors at runtime.

[0091] The business delivery module is responsible for resource scheduling based on the standard user resource demand description template, including preliminary screening of resources, matching of scheduling strategies and final screening of resources.

[0092] Based on the above technologies, corresponding plug-ins are developed and integrated into the computing network operating system, and a user-friendly front-end page is provided to facilitate users to automatically create computing resources through natural language input, thereby improving the usability of the computing network platform.

[0093] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling of computing power network resources based on large model intention perception, characterized in that: The steps include: S1: Build multiple intelligent agents with the same functions from different large model service platforms. Each intelligent agent consists of multiple nodes that implement different functions, including intent classification nodes, branch nodes, intent extraction conversion nodes, entity conversion nodes, feasibility analysis nodes, resource classification and template filling nodes, parameter checking nodes, and grammar checking nodes. S2: Design a prompt word template for each node. The prompt word template is a structured input format tailored for large language models to guide the model to generate expected outputs. S3: Build a standardized computing network resource requirement description template library, attribute semantic mapping table, and intent mapping table; S4: The resource type intent template library, attribute semantic mapping table and intent mapping table are stored in the knowledge base. The knowledge base converts the information into vectors for storage and builds an index structure to quickly retrieve information through a similarity algorithm. S5: Input the user's natural language description into the agent, and after being processed by the intent classification node, branch node, intent extraction conversion node, entity conversion node, feasibility analysis node, resource classification and template filling node, parameter check node and grammar check node, the final user resource demand description template is obtained and configured and issued; S6: During the configuration delivery process, the resources are initially screened according to the resource configuration information in the user resource requirement description template to obtain the candidate resource set T. The scheduling strategy is then matched according to the user resource requirement description template, and the candidate resource set T is further screened and sorted, and finally the appropriate resources are fed back to the user.

2. According to claim 1, the method for intelligent scheduling of computing power network resources based on large model intention perception is characterized in that: In step S1, each node is connected to multiple agents through multiple interfaces to enhance the fault tolerance of the system; The optimal agent is selected through the agent switching algorithm. The agent switching algorithm comprehensively considers three aspects: interface availability, response time and load balancing. The average response time of the interface is calculated by the sliding average method, the interface availability is checked regularly, and a comprehensive load indicator is introduced to quantify the load status of the interface, comprehensively considering the number of requests, processing time and resource utilization; the comprehensive load indicator adopts a dynamic attenuation mechanism to adapt to changes in system load. Finally, the score of each interface is calculated through a comprehensive evaluation indicator. When the interface is unavailable or the response time exceeds the threshold, the available interface with the highest comprehensive evaluation score is selected for switching to ensure the optimal performance of the system.

3. According to claim 2, the method for intelligent scheduling of computing power network resources based on large model intention perception is characterized in that: In step S2, the prompt word template decomposes a total task into multiple subtasks to guide the large language model to execute according to the specified rules, and continuously optimizes the prompt words according to the generated results to meet the current task requirements.

4. According to claim 3, the method for intelligent scheduling of computing power network resources based on large model intention perception is characterized in that: In step S3, the standard computing network resource demand description template library, i.e., the resource type intention template library, includes cloud host configuration information template, storage configuration information template, and network configuration information template, which are described in json format; The process of building a standardized computing network resource requirement description template library is as follows: Conduct demand analysis and research based on the computing network scheduling system to clarify resource allocation requirements; Design the template structure, define the standard attributes of each resource category, and determine the data type and value range; Write resource templates in JSON format and ensure that they comply with the syntax rules of the target platform through syntax checking and logic verification; The resource templates that have passed the verification are stored in the standard computing network resource requirement description template library.

5. According to claim 4, the method for intelligent scheduling of computing power network resources based on large model intention perception is characterized in that: In step S3, the attribute semantic mapping table contains the meaning of each field attribute in each standardized computing network resource requirement description template library under a specific computing network platform, so as to improve the understanding ability of the large language model to the standardized computing network resource requirement description template library; The intent mapping table contains as many expressions as possible for each field attribute value, realizing the mapping of user intent to computing network entities.

6. The method for intelligent scheduling of computing power network resources based on large model intention perception according to claim 5 is characterized in that: In step S4, the knowledge base parses the uploaded knowledge and slices the parsed knowledge to form knowledge units to facilitate the vectorization of knowledge; the knowledge units are vectorized using the Embedding vector model and converted into vector form to be stored in the vector database; an index structure is established for the vector database, and the index structure includes an inverted index, a KD tree or HNSW; and information is quickly retrieved through a similarity algorithm, and the similarity algorithm uses a cosine similarity algorithm.

7. The method for intelligent scheduling of computing power network resources based on large model intention perception according to claim 6 is characterized in that: In step S5, the intent classification node is completed based on the large language model, which is used to classify the user intent into two categories: general questions and answers and configuration information questions and answers according to the user's natural language description, and pass it to the corresponding process through the branch node. If it is a general question and answer, the large language model is directly used to answer the user's natural language description; if it is a configuration information question and answer, it is identified as a resource request and enters the subsequent intent extraction conversion node process; The intent extraction conversion node guides the task through the prompt word template, identifies the resource entities in the user's intent, including resource types such as cloud host, memory, network delay, etc., and converts the resource entities into JSON format; The entity conversion node further converts the resource type in the preliminary conversion result obtained by the attribute semantic mapping table and intent mapping table in the knowledge base and the intent extraction conversion node in the form of the computing network identifier expression to obtain the converted json format data, which is recorded as the computing network entity data.

8. The method for intelligent scheduling of computing power network resources based on large model intention perception according to claim 7 is characterized in that: In step S5, the feasibility analysis node determines whether the computing network platform has sufficient resources for allocation based on the output of the entity conversion node. The feasibility analysis node obtains the resource utilization of the computing network platform by adding the computing network resource query interface in the form of a plug-in; if any resource in the computing network platform cannot meet the resource configuration required by the user, the user's intention is considered infeasible and feedback is given to the user; otherwise, the user's intention is considered feasible and the next step is performed; When the user intention is feasible, the resource classification and template filling node classifies the resource type intention according to the resource type in the computing network entity data, matches the corresponding resource type intention template from the standardized computing network resource demand description template library in the knowledge base according to the classification result, and fills the matching resource type intention template library according to the data in the computing network entity data to obtain the user resource demand description template.

9. The method for intelligent scheduling of computing power network resources based on large model intention perception according to claim 8 is characterized in that: In step S5, the user resource requirement description template obtained by the resource classification and template filling node is subjected to feasibility check through the parameter check node and the syntax check node. After passing the check, the final user resource requirement description template is obtained.

10. The method for intelligent scheduling of computing power network resources based on large model intention perception according to claim 9 is characterized in that: The parameter check node compares the output of the resource classification and template filling node with the corresponding resource type intent template in the knowledge base, extracts the different contents between the two, that is, the content modified by the resource classification and template filling node to the resource type intent template, recorded as U1, and then compares U1 with the output of the entity conversion node. If they are the same, it is considered to meet the requirements, otherwise it does not meet the requirements, and the content of the resource classification and template filling node is modified according to the output of the entity conversion node, and output according to the modified content; According to the requirements of the computing network platform, grammar rules are pre-formulated. The grammar checking node performs grammar analysis on the generated user resource demand description template according to the pre-formulated grammar rules, checks whether the user resource demand description template generated by the intelligent agent has errors that violate grammar rules, and captures error information; If there is no syntax error, a final user resource requirement description template is generated; if there is a syntax error, the user resource requirement description template is modified according to the error information until there is no syntax error.

Images